Human-Guided Agentic AI for Skill Formation in Indian Higher Education: A Framework-Centred Study for Learner Modelling, Skill-Gap Diagnosis and Responsible Evaluation

Abstract

Artificial intelligence in education has a longer research history than the current wave of large language models. Intelligent tutoring systems, knowledge tracing, learning analytics and educational data mining established the methodological basis for adaptive support; generative AI subsequently made natural-language feedback, question generation, coding assistance and assessment redesign widely accessible; and agentic AI now promises multi-step workflows that can plan, retrieve, monitor and adapt. The difficulty is that agentic educational AI is advancing faster as a technology than as a validated classroom practice, and controlled evidence is especially scarce in Indian higher and engineering education, where scale, multilingual learners, Outcome-Based Education (OBE) requirements and data-protection duties shape what responsible deployment can look like. This paper addresses that gap through design-science reasoning: it synthesises foundational and recent evidence only to the extent needed to derive design requirements, and then develops SAKSHAM-AI, a human-guided agentic framework built on the principle of bounded autonomy. The framework integrates learner profiling, skill-gap diagnosis, knowledge tracing, retrieval-augmented content grounding, formative assessment, misconception-level feedback, self-regulated learning support, faculty dashboards, CO/PO outcome mapping and governance monitoring into a single auditable workflow in which agents recommend, retrieve, explain and flag while teachers retain authority over high-stakes academic decisions. We specify a trial-ready evaluation protocol-a 12–16-week pilot with a delayed-retention test-reporting knowledge gain, skill-gap closure, retention, independent transfer, equity, hallucination rate, explainability, privacy minimisation, escalation and OBE coverage. The contribution is therefore a bounded-autonomy, human-guided architecture together with its measurement and governance design, rather than a deployment claim; the framework is proposed and trial-ready but requires empirical validation in live institutional settings before any effectiveness claim can be made.